In the cutthroat AI world, Cerebras Systems, despite colossal wafers and bold claims, faces collapse. Its fragile foundation stems from brittle hardware, astonishingly narrow in practical application. Optimized primarily for LLaMA model finetuning, touted "speed boosts" often use speculative decoding, masking that many "supported" models run inefficiently, if at all. This rigidity mirrors a Formula 1 engine demanding one fuel under lab conditions.
This compounds with a poorly conceived software and kernel stack. The current infrastructure restricts Cerebras systems almost exclusively to LLaMA-family transformer LLMs. Modern vision pipelines, diffusion models, or other AI paradigms are beyond reach. In a rapidly evolving field, this narrow focus is a death sentence. A complete software overhaul is desperately needed, yet internal motivation is absent, hinting at leadership either oblivious or unwilling to invest in survival.\nTechnical woes are exacerbated by third-class ML expertise and rudderless, top-down leadership. Non-technical decision-makers lead to misguided strategies and poor adaptation. ML expertise is subpar and highly politicized; questionable ideas from figures like Head of AI Ganesh Venkatesh are reportedly prioritized, wasting resources.\nAn extreme resource shortage \u2013 Cerebras hardware and even GPUs \u2013 cripples meaningful work, making numbers in reports untrustworthy fiction. The chasm between rhetoric and reality widens.
A palpable resignation within the ranks further indicts the company. Employees are "waiting it out" or trading stocks, not contributing. This disengagement highlights deep problems, where personal finance trumps innovation.
Cerebras appears a bloated, incompetent, and culturally bankrupt company. Any stock price would be a scam. The rare "A players" have left or are leaving.
The writing is on the wall. Cerebras, despite grand pronouncements, is a company built on sand, destined to crumble under its limitations and poor choices.
Get out while you still can."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"Cerebras: The AI hardware giant facing imminent collapse"}},"_tags":["story","author_noob_hardy","story_45477162","ask_hn"],"author":"noob_hardy","children":[45478517,45479244,45480572,45486695,45510370],"created_at":"2025-10-04T22:08:08Z","created_at_i":1759615688,"num_comments":7,"objectID":"45477162","points":22,"story_id":45477162,"story_text":"Post by ex-employee.
In the cutthroat AI world, Cerebras Systems, despite colossal wafers and bold claims, faces collapse. Its fragile foundation stems from brittle hardware, astonishingly narrow in practical application. Optimized primarily for LLaMA model finetuning, touted "speed boosts" often use speculative decoding, masking that many "supported" models run inefficiently, if at all. This rigidity mirrors a Formula 1 engine demanding one fuel under lab conditions.
This compounds with a poorly conceived software and kernel stack. The current infrastructure restricts Cerebras systems almost exclusively to LLaMA-family transformer LLMs. Modern vision pipelines, diffusion models, or other AI paradigms are beyond reach. In a rapidly evolving field, this narrow focus is a death sentence. A complete software overhaul is desperately needed, yet internal motivation is absent, hinting at leadership either oblivious or unwilling to invest in survival.\nTechnical woes are exacerbated by third-class ML expertise and rudderless, top-down leadership. Non-technical decision-makers lead to misguided strategies and poor adaptation. ML expertise is subpar and highly politicized; questionable ideas from figures like Head of AI Ganesh Venkatesh are reportedly prioritized, wasting resources.\nAn extreme resource shortage \u2013 Cerebras hardware and even GPUs \u2013 cripples meaningful work, making numbers in reports untrustworthy fiction. The chasm between rhetoric and reality widens.
A palpable resignation within the ranks further indicts the company. Employees are "waiting it out" or trading stocks, not contributing. This disengagement highlights deep problems, where personal finance trumps innovation.
Cerebras appears a bloated, incompetent, and culturally bankrupt company. Any stock price would be a scam. The rare "A players" have left or are leaving.
The writing is on the wall. Cerebras, despite grand pronouncements, is a company built on sand, destined to crumble under its limitations and poor choices.
Get out while you still can.","title":"Cerebras: The AI hardware giant facing imminent collapse","updated_at":"2026-04-26T11:02:21Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"modeless"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"Cerebras runs Llama 3.1 70B at 2,100 tokens per second, live demo available"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://twitter.com/CerebrasSystems/status/1849467759517896955"}},"_tags":["story","author_modeless","story_41937581"],"author":"modeless","children":[41937596],"created_at":"2024-10-24T17:27:21Z","created_at_i":1729790841,"num_comments":0,"objectID":"41937581","points":6,"story_id":41937581,"title":"Cerebras runs Llama 3.1 70B at 2,100 tokens per second, live demo available","updated_at":"2024-10-25T19:02:30Z","url":"https://twitter.com/CerebrasSystems/status/1849467759517896955"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"montyanderson"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"Cerebras Inference"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://twitter.com/CerebrasSystems/status/1828464491677524311"}},"_tags":["story","author_montyanderson","story_41369685"],"author":"montyanderson","created_at":"2024-08-27T16:48:30Z","created_at_i":1724777310,"num_comments":0,"objectID":"41369685","points":4,"story_id":41369685,"title":"Cerebras Inference","updated_at":"2024-09-20T17:41:53Z","url":"https://twitter.com/CerebrasSystems/status/1828464491677524311"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"craftsquick"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"Cerebras releases open source GPT models"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.businesswire.com/news/home/20230328005366/en/Cerebras-Systems-Releases-Seven-New-GPT-Models-Trained-on-CS-2-Wafer-Scale-Systems"}},"_tags":["story","author_craftsquick","story_35376078"],"author":"craftsquick","created_at":"2023-03-30T17:17:33Z","created_at_i":1680196653,"num_comments":0,"objectID":"35376078","points":2,"story_id":35376078,"title":"Cerebras releases open source GPT models","updated_at":"2024-09-20T13:41:42Z","url":"https://www.businesswire.com/news/home/20230328005366/en/Cerebras-Systems-Releases-Seven-New-GPT-Models-Trained-on-CS-2-Wafer-Scale-Systems"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"HammadB"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Cerebras Releases Seven New GPT Models Trained on CS-2 Wafer-Scale System"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.businesswire.com/news/home/20230328005366/en/Cerebras-Systems-Releases-Seven-New-GPT-Models-Trained-on-CS-2-Wafer-Scale-Systems"}},"_tags":["story","author_HammadB","story_35347013"],"author":"HammadB","created_at":"2023-03-28T20:03:41Z","created_at_i":1680033821,"num_comments":0,"objectID":"35347013","points":2,"story_id":35347013,"title":"Cerebras Releases Seven New GPT Models Trained on CS-2 Wafer-Scale System","updated_at":"2024-09-20T13:38:33Z","url":"https://www.businesswire.com/news/home/20230328005366/en/Cerebras-Systems-Releases-Seven-New-GPT-Models-Trained-on-CS-2-Wafer-Scale-Systems"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"keveman"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"20B parameter GPT-3 model trained on a single Cerebras chip"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.cerebras.net/press-release/cerebras-systems-sets-record-for-largest-ai-models-ever-trained-on-a-single-device/"}},"_tags":["story","author_keveman","story_31836081"],"author":"keveman","created_at":"2022-06-22T14:01:02Z","created_at_i":1655906462,"num_comments":0,"objectID":"31836081","points":2,"story_id":31836081,"title":"20B parameter GPT-3 model trained on a single Cerebras chip","updated_at":"2024-09-20T11:25:09Z","url":"https://www.cerebras.net/press-release/cerebras-systems-sets-record-for-largest-ai-models-ever-trained-on-a-single-device/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"anandaverma18"},"title":{"matchLevel":"none","matchedWords":[],"value":"The Largest AI Chip with Trillion Transistors Unveiled"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://8bitwork.com/2019/08/25/cerebras-systems-unveiled-the-largest-ai-chip-with-trillion-transistor/"}},"_tags":["story","author_anandaverma18","story_20791407"],"author":"anandaverma18","created_at":"2019-08-25T04:45:17Z","created_at_i":1566708317,"num_comments":0,"objectID":"20791407","points":2,"story_id":20791407,"title":"The Largest AI Chip with Trillion Transistors Unveiled","updated_at":"2024-09-20T04:52:02Z","url":"https://8bitwork.com/2019/08/25/cerebras-systems-unveiled-the-largest-ai-chip-with-trillion-transistor/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"deepnotderp"},"title":{"matchLevel":"none","matchedWords":[],"value":"AI Chip Startup Worth Almost a Billion"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.forbes.com/sites/aarontilley/2017/08/31/ai-chip-cerebras-systems-investment/"}},"_tags":["story","author_deepnotderp","story_16518703"],"author":"deepnotderp","created_at":"2018-03-05T04:22:48Z","created_at_i":1520223768,"num_comments":0,"objectID":"16518703","points":2,"story_id":16518703,"title":"AI Chip Startup Worth Almost a Billion","updated_at":"2024-09-20T02:06:55Z","url":"https://www.forbes.com/sites/aarontilley/2017/08/31/ai-chip-cerebras-systems-investment/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rbanffy"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"Cerebras Announces 130x Improvement on Monte Carlo Particle Transport over A100"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.cerebras.net/press-release/cerebras-systems-announces-130x-performance-improvement-on-key-nuclear-energy-simulation-over-nvidia-a100-gpus"}},"_tags":["story","author_rbanffy","story_38261681"],"author":"rbanffy","created_at":"2023-11-14T10:54:15Z","created_at_i":1699959255,"num_comments":0,"objectID":"38261681","points":1,"story_id":38261681,"title":"Cerebras Announces 130x Improvement on Monte Carlo Particle Transport over A100","updated_at":"2024-09-20T15:39:37Z","url":"https://www.cerebras.net/press-release/cerebras-systems-announces-130x-performance-improvement-on-key-nuclear-energy-simulation-over-nvidia-a100-gpus"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"JoachimS"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Cerebras Extends AI System to \u2018Brain-Sized\u2019 Algorithms"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.eetimes.com/cerebras-extends-ai-system-to-brain-sized-algorithms/"}},"_tags":["story","author_JoachimS","story_28316019"],"author":"JoachimS","created_at":"2021-08-26T15:16:20Z","created_at_i":1629990980,"num_comments":0,"objectID":"28316019","points":2,"story_id":28316019,"title":"Cerebras Extends AI System to \u2018Brain-Sized\u2019 Algorithms","updated_at":"2024-09-20T09:14:04Z","url":"https://www.eetimes.com/cerebras-extends-ai-system-to-brain-sized-algorithms/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rbanffy"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Scaling the \"Memory Wall\" for Seismic Processing on Cerebras CS-2 Systems"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://dl.acm.org/doi/10.1145/3581784.3627042"}},"_tags":["story","author_rbanffy","story_39642553"],"author":"rbanffy","created_at":"2024-03-08T16:19:38Z","created_at_i":1709914778,"num_comments":0,"objectID":"39642553","points":1,"story_id":39642553,"title":"Scaling the \"Memory Wall\" for Seismic Processing on Cerebras CS-2 Systems","updated_at":"2024-09-20T16:36:56Z","url":"https://dl.acm.org/doi/10.1145/3581784.3627042"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ItsTotallyOn"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"World's Largest Chip Gets a New Home: Cerebras Launches CS-1 System"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.tomshardware.com/news/worlds-largest-chip-gets-a-new-home-cerebras-launches-cs-1-system"}},"_tags":["story","author_ItsTotallyOn","story_21574049"],"author":"ItsTotallyOn","created_at":"2019-11-19T15:42:57Z","created_at_i":1574178177,"num_comments":0,"objectID":"21574049","points":2,"story_id":21574049,"title":"World's Largest Chip Gets a New Home: Cerebras Launches CS-1 System","updated_at":"2024-09-20T05:12:25Z","url":"https://www.tomshardware.com/news/worlds-largest-chip-gets-a-new-home-cerebras-launches-cs-1-system"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Katydid"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Details on New 7nm Cerebras CS-2 Waferscale ML Training System"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.nextplatform.com/2021/04/20/one-giant-leap-for-waferscale-ai/"}},"_tags":["story","author_Katydid","story_26878213"],"author":"Katydid","created_at":"2021-04-20T18:00:17Z","created_at_i":1618941617,"num_comments":0,"objectID":"26878213","points":4,"story_id":26878213,"title":"Details on New 7nm Cerebras CS-2 Waferscale ML Training System","updated_at":"2024-09-20T08:26:19Z","url":"https://www.nextplatform.com/2021/04/20/one-giant-leap-for-waferscale-ai/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"smarvin2"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"After using Claude code for a few months, I set out to create my own version and realized that I spent most of the time creating a framework for organizing agent interactions. After a few more months of tinkering and playing around with ideas I created Wasmind.
Wasmind is a modular framework for building massively parallel agentic systems.
It's an actor based system where each actor is a wasm module. Actor's are composed together to create Agents and you can have 1-1000s of Agents running at once. Actors communicate through message passing where every message is broadcasted to every actor.
Wasmind can be used to build systems like Claude Code or really anything multi-agent you can dream of (examples included in GitHub: https://github.com/SilasMarvin/wasmind).
Wasmind solves a few key problems:\n1. Modular plug and play\n2. User-centered easy configuration\n3. User-defined and guaranteed enforceable safety and agent restrictions (coming soon)\n4. Allows easily composing any number of agents
You can configure it to use any LLM local or remote. I haven't tried qwen3-next but qwen3-coder (especially served by providers like Cerebras) has been incredibly fun to play with.
Thanks for checking my project out!"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["systems"],"value":"Show HN: Wasmind \u2013 A framework for building massively parallel agentic systems"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/SilasMarvin/wasmind"}},"_tags":["story","author_smarvin2","story_45232784","show_hn"],"author":"smarvin2","created_at":"2025-09-13T15:17:58Z","created_at_i":1757776678,"num_comments":0,"objectID":"45232784","points":3,"story_id":45232784,"story_text":"After using Claude code for a few months, I set out to create my own version and realized that I spent most of the time creating a framework for organizing agent interactions. After a few more months of tinkering and playing around with ideas I created Wasmind.
Wasmind is a modular framework for building massively parallel agentic systems.
It's an actor based system where each actor is a wasm module. Actor's are composed together to create Agents and you can have 1-1000s of Agents running at once. Actors communicate through message passing where every message is broadcasted to every actor.
Wasmind can be used to build systems like Claude Code or really anything multi-agent you can dream of (examples included in GitHub: https://github.com/SilasMarvin/wasmind).
Wasmind solves a few key problems:\n1. Modular plug and play\n2. User-centered easy configuration\n3. User-defined and guaranteed enforceable safety and agent restrictions (coming soon)\n4. Allows easily composing any number of agents
You can configure it to use any LLM local or remote. I haven't tried qwen3-next but qwen3-coder (especially served by providers like Cerebras) has been incredibly fun to play with.
Thanks for checking my project out!","title":"Show HN: Wasmind \u2013 A framework for building massively parallel agentic systems","updated_at":"2026-03-05T22:40:30Z","url":"https://github.com/SilasMarvin/wasmind"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"PaulHoule"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Cerebras Wafer-Scale Integration vs. Nvidia GPU-Based Systems for AI"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2503.11698"}},"_tags":["story","author_PaulHoule","story_43537705"],"author":"PaulHoule","created_at":"2025-03-31T17:48:00Z","created_at_i":1743443280,"num_comments":0,"objectID":"43537705","points":2,"story_id":43537705,"title":"Cerebras Wafer-Scale Integration vs. Nvidia GPU-Based Systems for AI","updated_at":"2025-03-31T19:25:32Z","url":"https://arxiv.org/abs/2503.11698"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"doener"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Cerebras 1.2T Chip Integrated with LLNL\u2019s Lassen System for AI Research"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://insidehpc.com/2020/08/cerebras-1-2-trillion-chip-integrated-with-llnls-lassen-system-for-ai-research/"}},"_tags":["story","author_doener","story_24249573"],"author":"doener","created_at":"2020-08-23T04:44:16Z","created_at_i":1598157856,"num_comments":0,"objectID":"24249573","points":2,"story_id":24249573,"title":"Cerebras 1.2T Chip Integrated with LLNL\u2019s Lassen System for AI Research","updated_at":"2024-09-20T06:47:46Z","url":"https://insidehpc.com/2020/08/cerebras-1-2-trillion-chip-integrated-with-llnls-lassen-system-for-ai-research/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rbanffy"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Cerebras Demonstrates Trillion Parameter Model Training on a Single CS-3 System"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://cerebras.ai/press-release/cerebras-demonstrates-trillion-parameter-model-training-on-a-single-cs-3-system"}},"_tags":["story","author_rbanffy","story_42408143"],"author":"rbanffy","created_at":"2024-12-13T13:14:50Z","created_at_i":1734095690,"num_comments":0,"objectID":"42408143","points":1,"story_id":42408143,"title":"Cerebras Demonstrates Trillion Parameter Model Training on a Single CS-3 System","updated_at":"2024-12-13T13:19:50Z","url":"https://cerebras.ai/press-release/cerebras-demonstrates-trillion-parameter-model-training-on-a-single-cs-3-system"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"galaxyLogic"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["systems"],"value":"Gargantuan computer system with 27M AI 'cores'"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["cerebras"],"value":"https://www.zdnet.com/article/ai-startup-cerebras-built-a-gargantuan-ai-computer-for-abu-dhabis-g42-with-27-million-ai-cores/"}},"_tags":["story","author_galaxyLogic","story_36835981"],"author":"galaxyLogic","created_at":"2023-07-23T15:01:44Z","created_at_i":1690124504,"num_comments":0,"objectID":"36835981","points":1,"story_id":36835981,"title":"Gargantuan computer system with 27M AI 'cores'","updated_at":"2024-09-20T14:36:49Z","url":"https://www.zdnet.com/article/ai-startup-cerebras-built-a-gargantuan-ai-computer-for-abu-dhabis-g42-with-27-million-ai-cores/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ttobbaybbob"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"I'm Bobby, CTO of Tako. We just launched our Knowledge Search API.
Our API takes natural-language prompts like "Nvidia M&A history" and returns visual answers and grounding text sourced from real-time, structured data (example: https://trytako.com/card/YHloo1Ea7GRnBr_s5r6s/).
Most AI systems can\u2019t effectively reason about real-time, structured data. One reason is access: a lot of the most valuable info is trapped in databases web crawlers can't index. Google solves this with a team of 2k+ engineers that ingest data (stocks, sports, etc) into a proprietary Knowledge Graph. Our goal is to offer developers the same knowledge search + visualization primitives Google has built, tailored for AI use cases, and delivered via API.
We seek to augment LLM\u2019s capabilities, and this means that most of our biggest technical challenges stem from not getting a lot for \u201cfree\u201d from LLMs. For example, RAG architectures that generate final outputs with LLMs introduce accuracy issues we can\u2019t tolerate, and are too slow. We\u2019ve built a Generative Augmented Search (GAS) architecture that uses LLMs (currently Llama 3.3-70B on Cerebras) to analyze input queries (~200 ms) but use deterministic retrieval for most output generation. The data in our knowledge graph generally isn\u2019t available in LLMs or the web, so we have to acquire it directly from sources (including licensing it from authoritative providers like S&P Global). A limitation of this approach is that some developers want us to offer the flexibility of LLM analysis across web sources, even if it means tolerating non-authoritative sourcing and some hallucination. We\u2019re working on some solutions to that now.
Curious to hear how other people are fighting hallucination.
I'd love your feedback on the product (and happy to discuss/answer questions about it/the tech stack)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Tako, a Knowledge Search API"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://trytako.com/playground/"}},"_tags":["story","author_ttobbaybbob","story_43986828","show_hn"],"author":"ttobbaybbob","children":[43987093,43987138,43987248,43987336,43988752],"created_at":"2025-05-14T17:04:56Z","created_at_i":1747242296,"num_comments":12,"objectID":"43986828","points":22,"story_id":43986828,"story_text":"I'm Bobby, CTO of Tako. We just launched our Knowledge Search API.
Our API takes natural-language prompts like "Nvidia M&A history" and returns visual answers and grounding text sourced from real-time, structured data (example: https://trytako.com/card/YHloo1Ea7GRnBr_s5r6s/).
Most AI systems can\u2019t effectively reason about real-time, structured data. One reason is access: a lot of the most valuable info is trapped in databases web crawlers can't index. Google solves this with a team of 2k+ engineers that ingest data (stocks, sports, etc) into a proprietary Knowledge Graph. Our goal is to offer developers the same knowledge search + visualization primitives Google has built, tailored for AI use cases, and delivered via API.
We seek to augment LLM\u2019s capabilities, and this means that most of our biggest technical challenges stem from not getting a lot for \u201cfree\u201d from LLMs. For example, RAG architectures that generate final outputs with LLMs introduce accuracy issues we can\u2019t tolerate, and are too slow. We\u2019ve built a Generative Augmented Search (GAS) architecture that uses LLMs (currently Llama 3.3-70B on Cerebras) to analyze input queries (~200 ms) but use deterministic retrieval for most output generation. The data in our knowledge graph generally isn\u2019t available in LLMs or the web, so we have to acquire it directly from sources (including licensing it from authoritative providers like S&P Global). A limitation of this approach is that some developers want us to offer the flexibility of LLM analysis across web sources, even if it means tolerating non-authoritative sourcing and some hallucination. We\u2019re working on some solutions to that now.
Curious to hear how other people are fighting hallucination.
I'd love your feedback on the product (and happy to discuss/answer questions about it/the tech stack)","title":"Show HN: Tako, a Knowledge Search API","updated_at":"2025-05-16T22:04:09Z","url":"https://trytako.com/playground/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"codelion"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"I've built an open-source implementation of Google DeepMind's AlphaEvolve system called OpenEvolve. It's an evolutionary coding agent that uses LLMs to discover and optimize algorithms through iterative evolution.
Try it out: https://github.com/codelion/openevolve
What is this?
OpenEvolve evolves entire codebases (not just single functions) by leveraging an ensemble of LLMs combined with automated evaluation. It follows the evolutionary approach described in the AlphaEvolve paper but is fully open source and configurable.
I built this because I wanted to experiment with evolutionary code generation and see if I could replicate DeepMind's results. The original system successfully improved Google's data centers and found new mathematical algorithms, but no implementation was released.
How it works:
The system has four main components that work together in an evolutionary loop:
1. Program Database: Stores programs and their metrics in a MAP-Elites inspired structure
2. Prompt Sampler: Creates context-rich prompts with past solutions
3. LLM Ensemble: Generates code modifications using multiple models
4. Evaluator Pool: Tests programs and provides feedback metrics
What you can do with it:
- Run existing examples to see evolution in action
- Define your own problems with custom evaluation functions
- Configure LLM backends (works with any OpenAI-compatible API)
- Use multiple LLMs in ensemble for better results
- Optimize algorithms with multiple objectives
Two examples I've replicated from the AlphaEvolve paper:
- Circle Packing: Evolved from simple geometric patterns to sophisticated mathematical optimization, reaching 99.97% of DeepMind's reported results (2.634 vs 2.635 sum of radii for n=26).
- Function Minimization: Transformed a random search into a complete simulated annealing algorithm with cooling schedules and adaptive step sizes.
Technical insights:
- Low latency LLMs are critical for rapid generation cycles
- Best results using Gemini-Flash-2.0-lite + Gemini-Flash-2.0 as the ensemble
- For the circle packing problem, Gemini-Flash-2.0 + Claude-Sonnet-3.7 performed best
- Cerebras AI's API provided the fastest inference speeds
- Two-phase approach (exploration then exploitation) worked best for complex problems
Getting started (takes < 2 minutes)
# Clone and install
git clone https://github.com/codelion/openevolve.git
cd openevolve
pip install -e .
# Run the function minimization example
python openevolve-run.py
examples/function_minimization/initial_program.py \\
examples/function_minimization/evaluator.py \\\n\n --config examples/function_minimization/config.yaml \\\n\n --iterations 50\nI'll be around to answer questions and discuss!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: OpenEvolve \u2013 open-source implementation of DeepMind's AlphaEvolve"}},"_tags":["story","author_codelion","story_44043625","show_hn"],"author":"codelion","children":[44046565],"created_at":"2025-05-20T16:57:44Z","created_at_i":1747760264,"num_comments":3,"objectID":"44043625","points":8,"story_id":44043625,"story_text":"I've built an open-source implementation of Google DeepMind's AlphaEvolve system called OpenEvolve. It's an evolutionary coding agent that uses LLMs to discover and optimize algorithms through iterative evolution.
Try it out: https://github.com/codelion/openevolve
What is this?
OpenEvolve evolves entire codebases (not just single functions) by leveraging an ensemble of LLMs combined with automated evaluation. It follows the evolutionary approach described in the AlphaEvolve paper but is fully open source and configurable.
I built this because I wanted to experiment with evolutionary code generation and see if I could replicate DeepMind's results. The original system successfully improved Google's data centers and found new mathematical algorithms, but no implementation was released.
How it works:
The system has four main components that work together in an evolutionary loop:
1. Program Database: Stores programs and their metrics in a MAP-Elites inspired structure
2. Prompt Sampler: Creates context-rich prompts with past solutions
3. LLM Ensemble: Generates code modifications using multiple models
4. Evaluator Pool: Tests programs and provides feedback metrics
What you can do with it:
- Run existing examples to see evolution in action
- Define your own problems with custom evaluation functions
- Configure LLM backends (works with any OpenAI-compatible API)
- Use multiple LLMs in ensemble for better results
- Optimize algorithms with multiple objectives
Two examples I've replicated from the AlphaEvolve paper:
- Circle Packing: Evolved from simple geometric patterns to sophisticated mathematical optimization, reaching 99.97% of DeepMind's reported results (2.634 vs 2.635 sum of radii for n=26).
- Function Minimization: Transformed a random search into a complete simulated annealing algorithm with cooling schedules and adaptive step sizes.
Technical insights:
- Low latency LLMs are critical for rapid generation cycles
- Best results using Gemini-Flash-2.0-lite + Gemini-Flash-2.0 as the ensemble
- For the circle packing problem, Gemini-Flash-2.0 + Claude-Sonnet-3.7 performed best
- Cerebras AI's API provided the fastest inference speeds
- Two-phase approach (exploration then exploitation) worked best for complex problems
Getting started (takes < 2 minutes)
# Clone and install
git clone https://github.com/codelion/openevolve.git
cd openevolve
pip install -e .
# Run the function minimization example
python openevolve-run.py
examples/function_minimization/initial_program.py \\
examples/function_minimization/evaluator.py \\\n\n --config examples/function_minimization/config.yaml \\\n\n --iterations 50\nI'll be around to answer questions and discuss!","title":"Show HN: OpenEvolve \u2013 open-source implementation of DeepMind's AlphaEvolve","updated_at":"2026-05-07T19:10:46Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"russ"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Hey HN, it\u2019s Russ - cofounder of LiveKit. An open source stack for building realtime AI applications.
We\u2019re sharing our first homegrown AI model for turn detection. Here\u2019s a live demo: https://cerebras.vercel.app/
Voice AI has come a long way in the last year. We now have end-to-end systems that can generate a response to user input in 300-500ms \u2014 human level speeds!
As latency reduces, a common problem that surfaces is the LLM responds too quickly. Any time there\u2019s a short pause in a user\u2019s speech, it ends up interrupting them. This is largely due to how voice AI applications perform \u201cturn detection\u201d \u2014 that is, figuring out when the user has finished speaking and when the model can run inference and respond.
Pretty much everyone uses a signal processing technique called voice activity detection (VAD). In a nutshell, it figures out when the audio signal switches from speech to silence and then triggers an end of turn once a configurable amount of silence has transpired.
One obvious delta between VAD and how humans do turn detection is we also consider the content of speech (i.e. what someone says). These past few months, we\u2019ve been working on an open weights, content-aware turn detection model for voice AI applications. It was fine-tuned from SmolLM v2 on text, runs on CPU (currently takes 50ms for inference), and uses speech transcriptions as input to predict when a user has completed a thought (also called an \u201cutterance\u201d). Since it was trained on text, notably it works well for pipeline-based architectures (i.e. STT \u21d2 LLM \u21d2 TTS).
We use this model together with VAD to make better predictions about whether a user is done speaking. Here\u2019s some demos --
- Podcast interview: https://youtu.be/EYDrSSEP0h0
- Ordering food: https://youtu.be/fcr8Y-3c4E0
- Providing shipping address: https://youtu.be/2pQWvd6xozw
- Customer support: https://youtu.be/YoSRg3ORKtQ
In our testing we\u2019ve found:
- 85% reduction in unintentional interruptions
- 3% false positives (where the user is done speaking, but the model thinks they aren\u2019t)
In practice, we still have work to do. We currently delay inference if the model predicts a < 15% chance the user is done speaking. This threshold misses a bunch of middle-of-the-pack probabilities.
Next steps are improving the model accuracy, tuning performance, and expanding to support more languages (only supports English rn). Separately, we\u2019re starting to explore an audio-based model that considers not just what someone says but how they say it, which can be used with natively multimodal models like GPT-4o that directly process and generate audio.
Code here: https://github.com/livekit/agents/tree/main/livekit-plugins/...
Let us know what you think!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Open-source turn detection model for voice AI"}},"_tags":["story","author_russ","story_42497868","show_hn"],"author":"russ","children":[42515545],"created_at":"2024-12-23T21:46:37Z","created_at_i":1734990397,"num_comments":1,"objectID":"42497868","points":8,"story_id":42497868,"story_text":"Hey HN, it\u2019s Russ - cofounder of LiveKit. An open source stack for building realtime AI applications.
We\u2019re sharing our first homegrown AI model for turn detection. Here\u2019s a live demo: https://cerebras.vercel.app/
Voice AI has come a long way in the last year. We now have end-to-end systems that can generate a response to user input in 300-500ms \u2014 human level speeds!
As latency reduces, a common problem that surfaces is the LLM responds too quickly. Any time there\u2019s a short pause in a user\u2019s speech, it ends up interrupting them. This is largely due to how voice AI applications perform \u201cturn detection\u201d \u2014 that is, figuring out when the user has finished speaking and when the model can run inference and respond.
Pretty much everyone uses a signal processing technique called voice activity detection (VAD). In a nutshell, it figures out when the audio signal switches from speech to silence and then triggers an end of turn once a configurable amount of silence has transpired.
One obvious delta between VAD and how humans do turn detection is we also consider the content of speech (i.e. what someone says). These past few months, we\u2019ve been working on an open weights, content-aware turn detection model for voice AI applications. It was fine-tuned from SmolLM v2 on text, runs on CPU (currently takes 50ms for inference), and uses speech transcriptions as input to predict when a user has completed a thought (also called an \u201cutterance\u201d). Since it was trained on text, notably it works well for pipeline-based architectures (i.e. STT \u21d2 LLM \u21d2 TTS).
We use this model together with VAD to make better predictions about whether a user is done speaking. Here\u2019s some demos --
- Podcast interview: https://youtu.be/EYDrSSEP0h0
- Ordering food: https://youtu.be/fcr8Y-3c4E0
- Providing shipping address: https://youtu.be/2pQWvd6xozw
- Customer support: https://youtu.be/YoSRg3ORKtQ
In our testing we\u2019ve found:
- 85% reduction in unintentional interruptions
- 3% false positives (where the user is done speaking, but the model thinks they aren\u2019t)
In practice, we still have work to do. We currently delay inference if the model predicts a < 15% chance the user is done speaking. This threshold misses a bunch of middle-of-the-pack probabilities.
Next steps are improving the model accuracy, tuning performance, and expanding to support more languages (only supports English rn). Separately, we\u2019re starting to explore an audio-based model that considers not just what someone says but how they say it, which can be used with natively multimodal models like GPT-4o that directly process and generate audio.
Code here: https://github.com/livekit/agents/tree/main/livekit-plugins/...
Let us know what you think!","title":"Show HN: Open-source turn detection model for voice AI","updated_at":"2025-08-13T00:43:33Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"theprophecy"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"I was tired of major job boards like LinkedIn/Indeed/Glass Door having spammy, promotions, and irrelevant job postings. Especially in ML where it was a pain to find MLE roles specializing in Computer Vision.
So I built Rocket Jobs. It looks at your resume and then uses semantic search to match you to relevant job postings based on your actual work experience. I spent a lot of time improving this "RAG" system by trying different embedding vendors and techniques to improve retrieval quality.
I intend to keep this 100% free. I'm able to run it for $0 rn. I'm also parsing 2M tokens a day all for free because I'm just using a 8B param llama for free rn from cerebras, groq, and together.ai. I plan to migrate to gemini later.
Also getting free postgres from aiven.io and free hosting from heroku student discount valid for 2 years
Link: https://www.rocketjobs.app/"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Web App that looks at your resume and matches you to jobs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.rocketjobs.app/"}},"_tags":["story","author_theprophecy","story_42120323","show_hn"],"author":"theprophecy","children":[42121980],"created_at":"2024-11-12T22:10:25Z","created_at_i":1731449425,"num_comments":7,"objectID":"42120323","points":4,"story_id":42120323,"story_text":"I was tired of major job boards like LinkedIn/Indeed/Glass Door having spammy, promotions, and irrelevant job postings. Especially in ML where it was a pain to find MLE roles specializing in Computer Vision.
So I built Rocket Jobs. It looks at your resume and then uses semantic search to match you to relevant job postings based on your actual work experience. I spent a lot of time improving this "RAG" system by trying different embedding vendors and techniques to improve retrieval quality.
I intend to keep this 100% free. I'm able to run it for $0 rn. I'm also parsing 2M tokens a day all for free because I'm just using a 8B param llama for free rn from cerebras, groq, and together.ai. I plan to migrate to gemini later.
Also getting free postgres from aiven.io and free hosting from heroku student discount valid for 2 years
Link: https://www.rocketjobs.app/","title":"Show HN: Web App that looks at your resume and matches you to jobs","updated_at":"2024-11-14T15:21:03Z","url":"https://www.rocketjobs.app/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"keatonlivermore"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"We've been running Cursor Ultra ($200/month) in Auto Mode for extended sessions \u2014 sometimes 8+ hours of autonomous development. This repository is the result: 170k+ lines of Zig, a complete RISC-V64 operating system.
*The Cursor Ultra Experience:*
Auto Mode changes everything. Instead of prompting line-by-line, we describe high-level goals:\n- "Fix all compilation errors in the JIT compiler"\n- "Add TCP socket support with proper error handling" \n- "Refactor this function to be under 64 lines"
The AI then works autonomously: reading files, understanding context, making coordinated changes across the codebase, writing tests, updating documentation. It's pair programming where your partner has read every file and never gets tired.
*What makes the code different:*
1. *AI-friendly style guide* \u2013 We created "Grain Style" specifically for AI-assisted development:\n - 64-line function maximum (2^6)\n - 128-character line maximum (2^7)\n - No recursion\n - 2+ assertions per function\n - Explicit types (u32/u64 never usize)\n - Every comment explains "why"
The AI follows these rules more consistently than humans would. 48 of 56 kernel files are fully compliant.\n3. *JIT Compiler* \u2013 Working x86_64 JIT that translates RISC-V instructions to native code. Near-native speed on x86 hosts.
4. *249 Tests* \u2013 Essential for AI-written code. The AI writes the tests too.
*Technical highlights:*
- Kernel boots on QEMU RISC-V64 with virt machine\n- 60+ syscalls (process, memory, IPC, network, audio, filesystem)\n- TCP/UDP networking with socket abstraction\n- Grainscript: a minimal scripting language with lexer, parser, interpreter\n- Process scheduler with priority queues\n- ELF loader supporting RISC-V64 binaries\n- Framebuffer graphics with dirty region tracking
*The economics:*
$200/month for Cursor Ultra. This codebase would have taken a small team several months at $15-20k+ in developer costs. We built it in weeks.
The catch: you still need to know what you're building. AI amplifies capability, it doesn't replace vision.
*Roadmap to Alpha:*
Bottleneck: Basin kernel \u2192 Vantage VM \u2192 Framework x86_64. Target application: *first-responder dispatch software*.
Contributors needed for ARM aarch64 (Apple Silicon) \u2014 Zig \u2192 C \u2192 Swift macOS app. Also designing Aurora (open-source iOS Cursor alternative) with two inference backends:\n- [Cursor CLI Ultra](https://cursor.com) ($200/month) \n- [Cerebras WSE](https://cerebras.ai) \u2014 spatial RAM, single-threaded bounded compute, deterministic latency
*Current status:*
Kernel boots, REPL works, Grainscript executes.
Built with Zig 0.15.2. Follows [Grain Style](docs/grain_style.md). MIT/Apache-2.0/BSD-3-Clause licensed."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Teamlibra/ry: a Zig framework for Cursor that makes prompting better"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://codeberg.org/teamlibra/ry"}},"_tags":["story","author_keatonlivermore","story_46666522","show_hn"],"author":"keatonlivermore","children":[46666561],"created_at":"2026-01-18T10:17:46Z","created_at_i":1768731466,"num_comments":1,"objectID":"46666522","points":2,"story_id":46666522,"story_text":"We've been running Cursor Ultra ($200/month) in Auto Mode for extended sessions \u2014 sometimes 8+ hours of autonomous development. This repository is the result: 170k+ lines of Zig, a complete RISC-V64 operating system.
*The Cursor Ultra Experience:*
Auto Mode changes everything. Instead of prompting line-by-line, we describe high-level goals:\n- "Fix all compilation errors in the JIT compiler"\n- "Add TCP socket support with proper error handling" \n- "Refactor this function to be under 64 lines"
The AI then works autonomously: reading files, understanding context, making coordinated changes across the codebase, writing tests, updating documentation. It's pair programming where your partner has read every file and never gets tired.
*What makes the code different:*
1. *AI-friendly style guide* \u2013 We created "Grain Style" specifically for AI-assisted development:\n - 64-line function maximum (2^6)\n - 128-character line maximum (2^7)\n - No recursion\n - 2+ assertions per function\n - Explicit types (u32/u64 never usize)\n - Every comment explains "why"
The AI follows these rules more consistently than humans would. 48 of 56 kernel files are fully compliant.\n3. *JIT Compiler* \u2013 Working x86_64 JIT that translates RISC-V instructions to native code. Near-native speed on x86 hosts.
4. *249 Tests* \u2013 Essential for AI-written code. The AI writes the tests too.
*Technical highlights:*
- Kernel boots on QEMU RISC-V64 with virt machine\n- 60+ syscalls (process, memory, IPC, network, audio, filesystem)\n- TCP/UDP networking with socket abstraction\n- Grainscript: a minimal scripting language with lexer, parser, interpreter\n- Process scheduler with priority queues\n- ELF loader supporting RISC-V64 binaries\n- Framebuffer graphics with dirty region tracking
*The economics:*
$200/month for Cursor Ultra. This codebase would have taken a small team several months at $15-20k+ in developer costs. We built it in weeks.
The catch: you still need to know what you're building. AI amplifies capability, it doesn't replace vision.
*Roadmap to Alpha:*
Bottleneck: Basin kernel \u2192 Vantage VM \u2192 Framework x86_64. Target application: *first-responder dispatch software*.
Contributors needed for ARM aarch64 (Apple Silicon) \u2014 Zig \u2192 C \u2192 Swift macOS app. Also designing Aurora (open-source iOS Cursor alternative) with two inference backends:\n- [Cursor CLI Ultra](https://cursor.com) ($200/month) \n- [Cerebras WSE](https://cerebras.ai) \u2014 spatial RAM, single-threaded bounded compute, deterministic latency
*Current status:*
Kernel boots, REPL works, Grainscript executes.
Built with Zig 0.15.2. Follows [Grain Style](docs/grain_style.md). MIT/Apache-2.0/BSD-3-Clause licensed.","title":"Show HN: Teamlibra/ry: a Zig framework for Cursor that makes prompting better","updated_at":"2026-03-05T23:27:15Z","url":"https://codeberg.org/teamlibra/ry"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tstockham"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"About two weeks ago, I posted Engram here, a memory layer for AI agents. The response was great and pushed me to keep building. Here's where things stand.
What changed since the last post:
DMR benchmark: 92.0% accuracy (460/500). Retrieval hit rate is 96.4%. This is competitive with systems backed by graph databases and Python ML stacks. Engram is TypeScript + SQLite.
LOCOMO benchmark (long-conversation memory): 80.0% across all 10 conversations, 1,540 questions. Full context scores 88.4% but costs 30x more tokens.
Bi-temporal memory model. Every memory has valid_from/valid_until timestamps. Point-in-time recall via asOf parameter. Contradiction detection automatically supersedes stale facts.
Hosted API launched on Fly.io with Stripe billing. Self-hosting remains free (bring your own Gemini key). Hosted tiers start at $29/mo.
OpenAI-compatible base URL. One env var to use Groq, Cerebras, Ollama, or any OpenAI-compatible provider instead of Gemini.
70 tests passing. Published engram-sdk@0.5.5 on npm.
What I learned:\nBenchmark scores are fragile. 13 commits to my core vault module dropped LOCOMO from 84.5% to 62%. I had to treat the eval suite like a regression test, run it after every meaningful change. If you're building a memory/RAG system and not doing this, you're flying blind.
The judge LLM matters more than you'd think. Switching from one model to another as the benchmark judge changed scores by 10+ points on the same data. Always disclose your judge model. We use Gemini 2.5 Flash.
Temporal context is everything. Memories without timestamps are almost useless for "when" questions. Prefixing memories with their conversation date and teaching the LLM to resolve relative dates ("yesterday," "last week") was the single biggest accuracy improvement.
The API is the product, not the SDK. 95% of users will hit a REST endpoint, not import a TypeScript module. I wish I'd built the hosted API sooner.
What's next: LangChain/CrewAI integrations, an Engram skill for OpenClaw agents, and getting the academic paper on arXiv.
Happy to answer questions about benchmarks, architecture, or the experience of building this as a PM who codes.\nGitHub: https://github.com/tstockham96/engram\nSite: https://engram.fyi\nnpm: https://www.npmjs.com/package/engram-sdk\nHosted API: https://engram-hosted.fly.dev"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Engram update \u2013 92% DMR, hosted API, lessons shipping agent memory"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/tstockham96/engram"}},"_tags":["story","author_tstockham","story_47249084","show_hn"],"author":"tstockham","created_at":"2026-03-04T15:39:49Z","created_at_i":1772638789,"num_comments":0,"objectID":"47249084","points":1,"story_id":47249084,"story_text":"About two weeks ago, I posted Engram here, a memory layer for AI agents. The response was great and pushed me to keep building. Here's where things stand.
What changed since the last post:
DMR benchmark: 92.0% accuracy (460/500). Retrieval hit rate is 96.4%. This is competitive with systems backed by graph databases and Python ML stacks. Engram is TypeScript + SQLite.
LOCOMO benchmark (long-conversation memory): 80.0% across all 10 conversations, 1,540 questions. Full context scores 88.4% but costs 30x more tokens.
Bi-temporal memory model. Every memory has valid_from/valid_until timestamps. Point-in-time recall via asOf parameter. Contradiction detection automatically supersedes stale facts.
Hosted API launched on Fly.io with Stripe billing. Self-hosting remains free (bring your own Gemini key). Hosted tiers start at $29/mo.
OpenAI-compatible base URL. One env var to use Groq, Cerebras, Ollama, or any OpenAI-compatible provider instead of Gemini.
70 tests passing. Published engram-sdk@0.5.5 on npm.
What I learned:\nBenchmark scores are fragile. 13 commits to my core vault module dropped LOCOMO from 84.5% to 62%. I had to treat the eval suite like a regression test, run it after every meaningful change. If you're building a memory/RAG system and not doing this, you're flying blind.
The judge LLM matters more than you'd think. Switching from one model to another as the benchmark judge changed scores by 10+ points on the same data. Always disclose your judge model. We use Gemini 2.5 Flash.
Temporal context is everything. Memories without timestamps are almost useless for "when" questions. Prefixing memories with their conversation date and teaching the LLM to resolve relative dates ("yesterday," "last week") was the single biggest accuracy improvement.
The API is the product, not the SDK. 95% of users will hit a REST endpoint, not import a TypeScript module. I wish I'd built the hosted API sooner.
What's next: LangChain/CrewAI integrations, an Engram skill for OpenClaw agents, and getting the academic paper on arXiv.
Happy to answer questions about benchmarks, architecture, or the experience of building this as a PM who codes.\nGitHub: https://github.com/tstockham96/engram\nSite: https://engram.fyi\nnpm: https://www.npmjs.com/package/engram-sdk\nHosted API: https://engram-hosted.fly.dev","title":"Show HN: Engram update \u2013 92% DMR, hosted API, lessons shipping agent memory","updated_at":"2026-03-05T23:41:31Z","url":"https://github.com/tstockham96/engram"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Mnexium"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Today we\u2019re open-sourcing the core memory engine behind Mnexium.com : CORE-MNX
GItHub (https://github.com/mnexium/core-mnx)\nNPM (https://www.npmjs.com/package/@mnexium/core)
For us, this is a product decision and a philosophy decision.
Memory infrastructure is becoming foundational for serious AI products, and we believe the core layer should be transparent, inspectable, and extensible by the teams building on top of it. We also just want feedback - we want to build the best memory system given the tools we have access to today. We also want to make LLMs perform better then they already do OOTB.
CORE-MNX is the backend layer that powers durable memory workflows:\n memory storage and retrieval,\n claim extraction and truth-state resolution,\n memory lifecycle management, and event streaming for real-time systems.\n It\u2019s Postgres-backed, API-first, and built to integrate into real production stacks.
We tried our best to make this system as standalone as possible. Ultimately, its fairly difficult we needed LLMs (Cerebras for fast token output, ChatGPT for intelligence etc), Databases for storage etc. We have intentionally made the project API interfaced so your project can be code agnostic.
Open-sourcing CORE lets builders:\n understand exactly how memory behavior works,\n self-host or extend the engine for their own products,\n and avoid reinventing the same memory infrastructure from scratch.
What stays on Mnexium.com\nMnexium\u2019s long-term direction is still the same: make AI systems more useful over time through durable memory and reliable recall. We've just figured out that hosting memory isnt the moat we once thought it was - the real moat we believe is making the LLM system(s) as easy to use as possible. The feature-set we've built around memory is what is differentiating.
Open-sourcing CORE is how we make that foundation available to everyone building in this space. Open to everyone to lend an opinion on improvements and how we make this problem solvable.
Would love feedback, opinions and bugs you may find. We release it isn't perfect, but certainly a good start we'd love to improve upon."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: OSS Durable Memory for LLMs"}},"_tags":["story","author_Mnexium","story_47088761","show_hn"],"author":"Mnexium","created_at":"2026-02-20T14:53:57Z","created_at_i":1771599237,"num_comments":0,"objectID":"47088761","points":1,"story_id":47088761,"story_text":"Today we\u2019re open-sourcing the core memory engine behind Mnexium.com : CORE-MNX
GItHub (https://github.com/mnexium/core-mnx)\nNPM (https://www.npmjs.com/package/@mnexium/core)
For us, this is a product decision and a philosophy decision.
Memory infrastructure is becoming foundational for serious AI products, and we believe the core layer should be transparent, inspectable, and extensible by the teams building on top of it. We also just want feedback - we want to build the best memory system given the tools we have access to today. We also want to make LLMs perform better then they already do OOTB.
CORE-MNX is the backend layer that powers durable memory workflows:\n memory storage and retrieval,\n claim extraction and truth-state resolution,\n memory lifecycle management, and event streaming for real-time systems.\n It\u2019s Postgres-backed, API-first, and built to integrate into real production stacks.
We tried our best to make this system as standalone as possible. Ultimately, its fairly difficult we needed LLMs (Cerebras for fast token output, ChatGPT for intelligence etc), Databases for storage etc. We have intentionally made the project API interfaced so your project can be code agnostic.
Open-sourcing CORE lets builders:\n understand exactly how memory behavior works,\n self-host or extend the engine for their own products,\n and avoid reinventing the same memory infrastructure from scratch.
What stays on Mnexium.com\nMnexium\u2019s long-term direction is still the same: make AI systems more useful over time through durable memory and reliable recall. We've just figured out that hosting memory isnt the moat we once thought it was - the real moat we believe is making the LLM system(s) as easy to use as possible. The feature-set we've built around memory is what is differentiating.
Open-sourcing CORE is how we make that foundation available to everyone building in this space. Open to everyone to lend an opinion on improvements and how we make this problem solvable.
Would love feedback, opinions and bugs you may find. We release it isn't perfect, but certainly a good start we'd love to improve upon.","title":"Show HN: OSS Durable Memory for LLMs","updated_at":"2026-03-05T23:36:25Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"accofrisk"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"In this post, I\u2019ll list all the parameters our AI currently measures for health monitoring. It doesn\u2019t matter which device the optical sensor is in (smartphone, stick, or box) and where the user places their wrist to measure the pulse wave. Today we use a watch form factor, but we don\u2019t want people to think our AI on the wrist is \u201cjust a watch.\u201d Suggest in the comments \u2013 what would you call it?
Cardiovascular System\n \u2013 Whole Blood Viscosity (WBV)\n \u2013 Fibrinogen (FIB)\n \u2013 Coronary Perfusion Pressure (CCP)\n \u2013 Myocardial Perfusion (CMBV)\n \u2013 Cardiac Output Force (CPO)\n \u2013 Myocardial Blood Supply and Consumption Rate (CMBR)\n \u2013 Right Ventricular Stroke Work (RVSW)\n \u2013 Myocardial Blood Demand (CMBN)\n \u2013 Cardiac Function Index (CFI)\n \u2013 Cardiac Index (CI)\n \u2013 Left Ventricular Stroke Work Index (LVWI)\n \u2013 Blood Pressure (SBP)\n \u2013 Heart Rate (HR)
Circulation and Microcirculation\n \u2013 Microcirculation Assessment\n \u2013 Microcirculation Half Update Rate (MHR)\n \u2013 Microcirculation Half Update Time (MRT)\n \u2013 Mean Retention Time of Microcirculation (MST)\n \u2013 Systemic Blood Flow (QS)
Lungs and Respiratory System\n \u2013 Arterial Oxygen Tension (PaO\u2082)\n \u2013 Vein Oxygen Saturation (SvO\u2082)\n \u2013 Pulmonary Vascular Resistance (PVR)\n \u2013 Artery Blood Oxygen Pressure / Oxygenation Index (OI)\n \u2013 Mean Pulmonary Arterial Pressure (MPAP)\n \u2013 Alveolar Ventilation (VA)\n \u2013 Oxygen Consumption (VO\u2082)\n \u2013 Pulmonary Vascular Permeability Index (PVPI)\n \u2013 Pulmonary Blood Volume (PBV)\n \u2013 Intrapulmonary Shunt Fraction (Qs/Qt)\n \u2013 Minute Ventilation (VE)\n \u2013 Arterial Partial Pressure of CO\u2082 (PaCO\u2082)\n \u2013 Vein Partial Pressure of CO\u2082 (PvCO\u2082)\n \u2013 Dead Space Ventilation (VD)\n \u2013 Rapid Shallow Breathing Index (RSBI)
Diabetes and Metabolism\n \u2013 Blood Glucose\n \u2013 Glycated Hemoglobin (HbA1c)\n \u2013 Fructosamine / Glycated Albumin\n \u2013 Insulin Resistance Index (HOMA-IR)\n \u2013 Beta-cell Function (HOMA-B)\n \u2013 Hypoglycemia-related Metrics\n \u2013 Carbohydrate, Fat, and Protein Intake\n \u2013 Metabolic Parameters (Lactate, Metabolism)
Liver\n \u2013 Blood Plasma Ammonia (BA)\n \u2013 Alanine Aminotransferase (ALT)\n \u2013 Aspartate Aminotransferase (AST)\n \u2013 AST/ALT Ratio\n \u2013 Lactate (Lac)\n \u2013 Indices: LAP, TyG, UHR, VAI
Kidneys\n \u2013 Glomerular Filtration Fraction (GFF)\n \u2013 Renal Perfusion Pressure (RPP)\n \u2013 Renal Blood Flow (RBF)\n \u2013 Uric Acid (UA)\n \u2013 Serum Creatinine (SCr)\n \u2013 Blood Urea Nitrogen (BUN)
Brain and Nervous System\n \u2013 Cerebral Blood Flow Resistance (CVR)\n \u2013 Jugular Oxygenation (SjVO\u2082)\n \u2013 Cerebral Perfusion Pressure (CPP)\n \u2013 Jugular Venous Oxygen Content (CjVO\u2082)\n \u2013 Cerebral Blood Flow (CBF)\n \u2013 Cerebral Blood Volume (CBV)\n \u2013 Intracranial Pressure (ICP)\n \u2013 Carotid-Venous Oxygen Content (AjVDo\u2082)\n \u2013 Cerebral Oxygen Metabolic Rate (CMRO\u2082)
Mental Health\n \u2013 Anxiety\n \u2013 Depression\n \u2013 Relaxation\n \u2013 Fear\n \u2013 Positivity\n \u2013 Anger\n \u2013 Stress\n \u2013 Sadness\n \u2013 Mania
Sleep\n \u2013 Apnea\n \u2013 Snoring\n \u2013 Sleep Duration\n \u2013 Sleep Phases
Nutrition and Physical Activity\n \u2013 Body Mass Index (BMI)\n \u2013 Resting Metabolic Rate (RMR)\n \u2013 Fat Percentage\n \u2013 Muscle Mass (%)\n \u2013 Obese Mass\n \u2013 Daily Calorie Intake\n \u2013 Carbohydrate, Fat, and Protein Intake
Female Health Parameters\n \u2013 Menstrual Phase\n \u2013 Follicular Phase\n \u2013 Predicted Ovulation\n \u2013 Luteal Phase\n \u2013 Body Mass Index (BMI)\n \u2013 Psychological State \u2013 rating scale\n \u2013 Anemia \u2013 present/absent\n \u2013 Fatigue \u2013 rating scale\n \u2013 Breathing Rate \u2013 breaths per minute\n \u2013 Risk of Premature Heartbeats (Extrasystoles) \u2013 present/rating\n \u2013 Fibrinogen \u2013 grams per liter\n \u2013 Hemoglobin \u2013 grams per deciliter
Child Health Parameters (ages 5\u201317) - Physical Health\n \u2013 Body Temperature\n \u2013 Blood Oxygen Saturation\n \u2013 Heart Rate\n \u2013 Level of Physical Activity\n \u2013 Blood Pressure\n \u2013 Sleep Quality and Sleep Phases
Child Health Parameters (ages 5\u201317) - Mental Health\n \u2013 Academic Pressure\n \u2013 Interpersonal Relationships\n \u2013 Physical and Mental Health Balance\n \u2013 Emotional Background\n \u2013 Behavioral Habits\n \u2013 Personality Traits"},"title":{"matchLevel":"none","matchedWords":[],"value":"Goodbye Smartwatches, Hello Health AI on Your Wrist"}},"_tags":["story","author_accofrisk","story_46896833","ask_hn"],"author":"accofrisk","children":[46897179],"created_at":"2026-02-05T07:37:35Z","created_at_i":1770277055,"num_comments":1,"objectID":"46896833","points":1,"story_id":46896833,"story_text":"In this post, I\u2019ll list all the parameters our AI currently measures for health monitoring. It doesn\u2019t matter which device the optical sensor is in (smartphone, stick, or box) and where the user places their wrist to measure the pulse wave. Today we use a watch form factor, but we don\u2019t want people to think our AI on the wrist is \u201cjust a watch.\u201d Suggest in the comments \u2013 what would you call it?
Cardiovascular System\n \u2013 Whole Blood Viscosity (WBV)\n \u2013 Fibrinogen (FIB)\n \u2013 Coronary Perfusion Pressure (CCP)\n \u2013 Myocardial Perfusion (CMBV)\n \u2013 Cardiac Output Force (CPO)\n \u2013 Myocardial Blood Supply and Consumption Rate (CMBR)\n \u2013 Right Ventricular Stroke Work (RVSW)\n \u2013 Myocardial Blood Demand (CMBN)\n \u2013 Cardiac Function Index (CFI)\n \u2013 Cardiac Index (CI)\n \u2013 Left Ventricular Stroke Work Index (LVWI)\n \u2013 Blood Pressure (SBP)\n \u2013 Heart Rate (HR)
Circulation and Microcirculation\n \u2013 Microcirculation Assessment\n \u2013 Microcirculation Half Update Rate (MHR)\n \u2013 Microcirculation Half Update Time (MRT)\n \u2013 Mean Retention Time of Microcirculation (MST)\n \u2013 Systemic Blood Flow (QS)
Lungs and Respiratory System\n \u2013 Arterial Oxygen Tension (PaO\u2082)\n \u2013 Vein Oxygen Saturation (SvO\u2082)\n \u2013 Pulmonary Vascular Resistance (PVR)\n \u2013 Artery Blood Oxygen Pressure / Oxygenation Index (OI)\n \u2013 Mean Pulmonary Arterial Pressure (MPAP)\n \u2013 Alveolar Ventilation (VA)\n \u2013 Oxygen Consumption (VO\u2082)\n \u2013 Pulmonary Vascular Permeability Index (PVPI)\n \u2013 Pulmonary Blood Volume (PBV)\n \u2013 Intrapulmonary Shunt Fraction (Qs/Qt)\n \u2013 Minute Ventilation (VE)\n \u2013 Arterial Partial Pressure of CO\u2082 (PaCO\u2082)\n \u2013 Vein Partial Pressure of CO\u2082 (PvCO\u2082)\n \u2013 Dead Space Ventilation (VD)\n \u2013 Rapid Shallow Breathing Index (RSBI)
Diabetes and Metabolism\n \u2013 Blood Glucose\n \u2013 Glycated Hemoglobin (HbA1c)\n \u2013 Fructosamine / Glycated Albumin\n \u2013 Insulin Resistance Index (HOMA-IR)\n \u2013 Beta-cell Function (HOMA-B)\n \u2013 Hypoglycemia-related Metrics\n \u2013 Carbohydrate, Fat, and Protein Intake\n \u2013 Metabolic Parameters (Lactate, Metabolism)
Liver\n \u2013 Blood Plasma Ammonia (BA)\n \u2013 Alanine Aminotransferase (ALT)\n \u2013 Aspartate Aminotransferase (AST)\n \u2013 AST/ALT Ratio\n \u2013 Lactate (Lac)\n \u2013 Indices: LAP, TyG, UHR, VAI
Kidneys\n \u2013 Glomerular Filtration Fraction (GFF)\n \u2013 Renal Perfusion Pressure (RPP)\n \u2013 Renal Blood Flow (RBF)\n \u2013 Uric Acid (UA)\n \u2013 Serum Creatinine (SCr)\n \u2013 Blood Urea Nitrogen (BUN)
Brain and Nervous System\n \u2013 Cerebral Blood Flow Resistance (CVR)\n \u2013 Jugular Oxygenation (SjVO\u2082)\n \u2013 Cerebral Perfusion Pressure (CPP)\n \u2013 Jugular Venous Oxygen Content (CjVO\u2082)\n \u2013 Cerebral Blood Flow (CBF)\n \u2013 Cerebral Blood Volume (CBV)\n \u2013 Intracranial Pressure (ICP)\n \u2013 Carotid-Venous Oxygen Content (AjVDo\u2082)\n \u2013 Cerebral Oxygen Metabolic Rate (CMRO\u2082)
Mental Health\n \u2013 Anxiety\n \u2013 Depression\n \u2013 Relaxation\n \u2013 Fear\n \u2013 Positivity\n \u2013 Anger\n \u2013 Stress\n \u2013 Sadness\n \u2013 Mania
Sleep\n \u2013 Apnea\n \u2013 Snoring\n \u2013 Sleep Duration\n \u2013 Sleep Phases
Nutrition and Physical Activity\n \u2013 Body Mass Index (BMI)\n \u2013 Resting Metabolic Rate (RMR)\n \u2013 Fat Percentage\n \u2013 Muscle Mass (%)\n \u2013 Obese Mass\n \u2013 Daily Calorie Intake\n \u2013 Carbohydrate, Fat, and Protein Intake
Female Health Parameters\n \u2013 Menstrual Phase\n \u2013 Follicular Phase\n \u2013 Predicted Ovulation\n \u2013 Luteal Phase\n \u2013 Body Mass Index (BMI)\n \u2013 Psychological State \u2013 rating scale\n \u2013 Anemia \u2013 present/absent\n \u2013 Fatigue \u2013 rating scale\n \u2013 Breathing Rate \u2013 breaths per minute\n \u2013 Risk of Premature Heartbeats (Extrasystoles) \u2013 present/rating\n \u2013 Fibrinogen \u2013 grams per liter\n \u2013 Hemoglobin \u2013 grams per deciliter
Child Health Parameters (ages 5\u201317) - Physical Health\n \u2013 Body Temperature\n \u2013 Blood Oxygen Saturation\n \u2013 Heart Rate\n \u2013 Level of Physical Activity\n \u2013 Blood Pressure\n \u2013 Sleep Quality and Sleep Phases
Child Health Parameters (ages 5\u201317) - Mental Health\n \u2013 Academic Pressure\n \u2013 Interpersonal Relationships\n \u2013 Physical and Mental Health Balance\n \u2013 Emotional Background\n \u2013 Behavioral Habits\n \u2013 Personality Traits","title":"Goodbye Smartwatches, Hello Health AI on Your Wrist","updated_at":"2026-03-05T23:32:06Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rajnathani"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"Cerebral Organoids Flunk Comparison to Developing Nervous System"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"https://www.ucsf.edu/news/2020/01/416526/not-brains-dish-cerebral-organoids-flunk-comparison-developing-nervous-system"}},"_tags":["story","author_rajnathani","story_22190732"],"author":"rajnathani","created_at":"2020-01-30T11:58:15Z","created_at_i":1580385495,"num_comments":0,"objectID":"22190732","points":1,"story_id":22190732,"title":"Cerebral Organoids Flunk Comparison to Developing Nervous System","updated_at":"2024-09-20T05:35:41Z","url":"https://www.ucsf.edu/news/2020/01/416526/not-brains-dish-cerebral-organoids-flunk-comparison-developing-nervous-system"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"accofrisk"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["cerebras","systems"],"value":"For the past decade, wearable health devices have been limited to measuring 20 or so metrics - heart rate, blood pressure, blood oxygen saturation, and steps. Steps, in particular, have become a kind of health ritual: walk more, count them carefully, and all will be fine. Smartwatches now offer dozens of sports modes, but the core value hasn\u2019t changed.
Despite record-breaking sales, wearable devices still don\u2019t provide meaningful health insight. What are users really paying for, and what does a subscription deliver? The real breakthrough comes from turning a wearable into a full laboratory on the wrist - and AI is already making this possible.
By early 2026, our team trained AI to measure more than 300 health parameters from a single wrist pulse signal. The data is sent to an AI-powered cloud, where it is transformed into a full health profile. Users access insights across metabolism and endocrinology, blood and microcirculation, cardiovascular and respiratory function, liver and kidney health, brain and cerebral circulation, as well as lifestyle factors like sleep, activity, stress, emotions, nutrition, and medication adherence. The result is a system-level view rather than isolated numbers.
Continuous measurement during sleep is particularly valuable. Nighttime data is cleaner, more stable, and far more predictive than daytime readings. Our \u201cDigital Sleep\u201d mode enables early risk detection and long-term trend analysis that would otherwise be impossible.
We go beyond tracking. Abnormal readings already signal potential problems, but our AI predicts disease risks before symptoms appear. The Health Assistant evaluates risks, connects physiological data with lifestyle patterns, and generates personalized recommendations. The most important advice remains - consult a physician and pursue further evaluation when necessary.
Our sensor architecture is simple - devices only need to collect a pulse signal. This allows integration into third-party devices and large-scale remote health monitoring. Organizations can track employee or patient health, improve safety, and make decisions based on real physiological data. Enterprise use cases include remote medical screenings, restricting vehicle operation if a driver is unwell, or temporarily removing an employee from duty due to stress - even when no external signs are visible.
The key transformation in health monitoring is the shift from measurement to prediction. AI turns wearables from fitness gadgets into preventive and predictive health tools. Cloud intelligence makes any device a full monitoring platform, while sleep-based continuous data provides the most reliable foundation for long-term risk assessment.
Soon, the first question a doctor asks may no longer be \u201cWhat are your symptoms?\u201d but \u201cPlease share your health data from the past week.\u201d Don\u2019t explain - just share."},"title":{"matchLevel":"none","matchedWords":[],"value":"Hack Your Health and Get 300 Health Metrics with AI"}},"_tags":["story","author_accofrisk","story_46846392","ask_hn"],"author":"accofrisk","children":[46846657],"created_at":"2026-02-01T14:19:44Z","created_at_i":1769955584,"num_comments":4,"objectID":"46846392","points":1,"story_id":46846392,"story_text":"For the past decade, wearable health devices have been limited to measuring 20 or so metrics - heart rate, blood pressure, blood oxygen saturation, and steps. Steps, in particular, have become a kind of health ritual: walk more, count them carefully, and all will be fine. Smartwatches now offer dozens of sports modes, but the core value hasn\u2019t changed.
Despite record-breaking sales, wearable devices still don\u2019t provide meaningful health insight. What are users really paying for, and what does a subscription deliver? The real breakthrough comes from turning a wearable into a full laboratory on the wrist - and AI is already making this possible.
By early 2026, our team trained AI to measure more than 300 health parameters from a single wrist pulse signal. The data is sent to an AI-powered cloud, where it is transformed into a full health profile. Users access insights across metabolism and endocrinology, blood and microcirculation, cardiovascular and respiratory function, liver and kidney health, brain and cerebral circulation, as well as lifestyle factors like sleep, activity, stress, emotions, nutrition, and medication adherence. The result is a system-level view rather than isolated numbers.
Continuous measurement during sleep is particularly valuable. Nighttime data is cleaner, more stable, and far more predictive than daytime readings. Our \u201cDigital Sleep\u201d mode enables early risk detection and long-term trend analysis that would otherwise be impossible.
We go beyond tracking. Abnormal readings already signal potential problems, but our AI predicts disease risks before symptoms appear. The Health Assistant evaluates risks, connects physiological data with lifestyle patterns, and generates personalized recommendations. The most important advice remains - consult a physician and pursue further evaluation when necessary.
Our sensor architecture is simple - devices only need to collect a pulse signal. This allows integration into third-party devices and large-scale remote health monitoring. Organizations can track employee or patient health, improve safety, and make decisions based on real physiological data. Enterprise use cases include remote medical screenings, restricting vehicle operation if a driver is unwell, or temporarily removing an employee from duty due to stress - even when no external signs are visible.
The key transformation in health monitoring is the shift from measurement to prediction. AI turns wearables from fitness gadgets into preventive and predictive health tools. Cloud intelligence makes any device a full monitoring platform, while sleep-based continuous data provides the most reliable foundation for long-term risk assessment.
Soon, the first question a doctor asks may no longer be \u201cWhat are your symptoms?\u201d but \u201cPlease share your health data from the past week.\u201d Don\u2019t explain - just share.","title":"Hack Your Health and Get 300 Health Metrics with AI","updated_at":"2026-03-05T23:28:35Z"}],"hitsPerPage":50,"nbHits":40,"nbPages":1,"page":0,"params":"query=Cerebras+Systems&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":32,"processingTimingsMS":{"_request":{"roundTrip":18},"afterFetch":{"format":{"highlighting":3,"total":4}},"fetch":{"query":12,"scanning":18,"total":31},"total":33},"query":"Cerebras Systems","serverTimeMS":39}